1/*
  2 * Sleepable Read-Copy Update mechanism for mutual exclusion.
  3 *
  4 * This program is free software; you can redistribute it and/or modify
  5 * it under the terms of the GNU General Public License as published by
  6 * the Free Software Foundation; either version 2 of the License, or
  7 * (at your option) any later version.
  8 *
  9 * This program is distributed in the hope that it will be useful,
 10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 12 * GNU General Public License for more details.
 13 *
 14 * You should have received a copy of the GNU General Public License
 15 * along with this program; if not, write to the Free Software
 16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
 17 *
 18 * Copyright (C) IBM Corporation, 2006
 19 *
 20 * Author: Paul McKenney <paulmck@us.ibm.com>
 21 *
 22 * For detailed explanation of Read-Copy Update mechanism see -
 23 * 		Documentation/RCU/ *.txt
 24 *
 25 */
 26
 27#include <linux/export.h>
 28#include <linux/mutex.h>
 29#include <linux/percpu.h>
 30#include <linux/preempt.h>
 31#include <linux/rcupdate.h>
 32#include <linux/sched.h>
 33#include <linux/smp.h>
 34#include <linux/delay.h>
 35#include <linux/srcu.h>
 36
 37/*
 38 * Initialize an rcu_batch structure to empty.
 39 */
 40static inline void rcu_batch_init(struct rcu_batch *b)
 41{
 42	b->head = NULL;
 43	b->tail = &b->head;
 44}
 45
 46/*
 47 * Enqueue a callback onto the tail of the specified rcu_batch structure.
 48 */
 49static inline void rcu_batch_queue(struct rcu_batch *b, struct rcu_head *head)
 50{
 51	*b->tail = head;
 52	b->tail = &head->next;
 53}
 54
 55/*
 56 * Is the specified rcu_batch structure empty?
 57 */
 58static inline bool rcu_batch_empty(struct rcu_batch *b)
 59{
 60	return b->tail == &b->head;
 61}
 62
 63/*
 64 * Remove the callback at the head of the specified rcu_batch structure
 65 * and return a pointer to it, or return NULL if the structure is empty.
 66 */
 67static inline struct rcu_head *rcu_batch_dequeue(struct rcu_batch *b)
 68{
 69	struct rcu_head *head;
 70
 71	if (rcu_batch_empty(b))
 72		return NULL;
 73
 74	head = b->head;
 75	b->head = head->next;
 76	if (b->tail == &head->next)
 77		rcu_batch_init(b);
 78
 79	return head;
 80}
 81
 82/*
 83 * Move all callbacks from the rcu_batch structure specified by "from" to
 84 * the structure specified by "to".
 85 */
 86static inline void rcu_batch_move(struct rcu_batch *to, struct rcu_batch *from)
 87{
 88	if (!rcu_batch_empty(from)) {
 89		*to->tail = from->head;
 90		to->tail = from->tail;
 91		rcu_batch_init(from);
 92	}
 93}
 94
 95/* single-thread state-machine */
 96static void process_srcu(struct work_struct *work);
 97
 98static int init_srcu_struct_fields(struct srcu_struct *sp)
 99{
100	sp->completed = 0;
101	spin_lock_init(&sp->queue_lock);
102	sp->running = false;
103	rcu_batch_init(&sp->batch_queue);
104	rcu_batch_init(&sp->batch_check0);
105	rcu_batch_init(&sp->batch_check1);
106	rcu_batch_init(&sp->batch_done);
107	INIT_DELAYED_WORK(&sp->work, process_srcu);
108	sp->per_cpu_ref = alloc_percpu(struct srcu_struct_array);
109	return sp->per_cpu_ref ? 0 : -ENOMEM;
110}
111
112#ifdef CONFIG_DEBUG_LOCK_ALLOC
113
114int __init_srcu_struct(struct srcu_struct *sp, const char *name,
115		       struct lock_class_key *key)
116{
117	/* Don't re-initialize a lock while it is held. */
118	debug_check_no_locks_freed((void *)sp, sizeof(*sp));
119	lockdep_init_map(&sp->dep_map, name, key, 0);
120	return init_srcu_struct_fields(sp);
121}
122EXPORT_SYMBOL_GPL(__init_srcu_struct);
123
124#else /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
125
126/**
127 * init_srcu_struct - initialize a sleep-RCU structure
128 * @sp: structure to initialize.
129 *
130 * Must invoke this on a given srcu_struct before passing that srcu_struct
131 * to any other function.  Each srcu_struct represents a separate domain
132 * of SRCU protection.
133 */
134int init_srcu_struct(struct srcu_struct *sp)
135{
136	return init_srcu_struct_fields(sp);
137}
138EXPORT_SYMBOL_GPL(init_srcu_struct);
139
140#endif /* #else #ifdef CONFIG_DEBUG_LOCK_ALLOC */
141
142/*
143 * Returns approximate total of the readers' ->seq[] values for the
144 * rank of per-CPU counters specified by idx.
145 */
146static unsigned long srcu_readers_seq_idx(struct srcu_struct *sp, int idx)
147{
148	int cpu;
149	unsigned long sum = 0;
150	unsigned long t;
151
152	for_each_possible_cpu(cpu) {
153		t = ACCESS_ONCE(per_cpu_ptr(sp->per_cpu_ref, cpu)->seq[idx]);
154		sum += t;
155	}
156	return sum;
157}
158
159/*
160 * Returns approximate number of readers active on the specified rank
161 * of the per-CPU ->c[] counters.
162 */
163static unsigned long srcu_readers_active_idx(struct srcu_struct *sp, int idx)
164{
165	int cpu;
166	unsigned long sum = 0;
167	unsigned long t;
168
169	for_each_possible_cpu(cpu) {
170		t = ACCESS_ONCE(per_cpu_ptr(sp->per_cpu_ref, cpu)->c[idx]);
171		sum += t;
172	}
173	return sum;
174}
175
176/*
177 * Return true if the number of pre-existing readers is determined to
178 * be stably zero.  An example unstable zero can occur if the call
179 * to srcu_readers_active_idx() misses an __srcu_read_lock() increment,
180 * but due to task migration, sees the corresponding __srcu_read_unlock()
181 * decrement.  This can happen because srcu_readers_active_idx() takes
182 * time to sum the array, and might in fact be interrupted or preempted
183 * partway through the summation.
184 */
185static bool srcu_readers_active_idx_check(struct srcu_struct *sp, int idx)
186{
187	unsigned long seq;
188
189	seq = srcu_readers_seq_idx(sp, idx);
190
191	/*
192	 * The following smp_mb() A pairs with the smp_mb() B located in
193	 * __srcu_read_lock().  This pairing ensures that if an
194	 * __srcu_read_lock() increments its counter after the summation
195	 * in srcu_readers_active_idx(), then the corresponding SRCU read-side
196	 * critical section will see any changes made prior to the start
197	 * of the current SRCU grace period.
198	 *
199	 * Also, if the above call to srcu_readers_seq_idx() saw the
200	 * increment of ->seq[], then the call to srcu_readers_active_idx()
201	 * must see the increment of ->c[].
202	 */
203	smp_mb(); /* A */
204
205	/*
206	 * Note that srcu_readers_active_idx() can incorrectly return
207	 * zero even though there is a pre-existing reader throughout.
208	 * To see this, suppose that task A is in a very long SRCU
209	 * read-side critical section that started on CPU 0, and that
210	 * no other reader exists, so that the sum of the counters
211	 * is equal to one.  Then suppose that task B starts executing
212	 * srcu_readers_active_idx(), summing up to CPU 1, and then that
213	 * task C starts reading on CPU 0, so that its increment is not
214	 * summed, but finishes reading on CPU 2, so that its decrement
215	 * -is- summed.  Then when task B completes its sum, it will
216	 * incorrectly get zero, despite the fact that task A has been
217	 * in its SRCU read-side critical section the whole time.
218	 *
219	 * We therefore do a validation step should srcu_readers_active_idx()
220	 * return zero.
221	 */
222	if (srcu_readers_active_idx(sp, idx) != 0)
223		return false;
224
225	/*
226	 * The remainder of this function is the validation step.
227	 * The following smp_mb() D pairs with the smp_mb() C in
228	 * __srcu_read_unlock().  If the __srcu_read_unlock() was seen
229	 * by srcu_readers_active_idx() above, then any destructive
230	 * operation performed after the grace period will happen after
231	 * the corresponding SRCU read-side critical section.
232	 *
233	 * Note that there can be at most NR_CPUS worth of readers using
234	 * the old index, which is not enough to overflow even a 32-bit
235	 * integer.  (Yes, this does mean that systems having more than
236	 * a billion or so CPUs need to be 64-bit systems.)  Therefore,
237	 * the sum of the ->seq[] counters cannot possibly overflow.
238	 * Therefore, the only way that the return values of the two
239	 * calls to srcu_readers_seq_idx() can be equal is if there were
240	 * no increments of the corresponding rank of ->seq[] counts
241	 * in the interim.  But the missed-increment scenario laid out
242	 * above includes an increment of the ->seq[] counter by
243	 * the corresponding __srcu_read_lock().  Therefore, if this
244	 * scenario occurs, the return values from the two calls to
245	 * srcu_readers_seq_idx() will differ, and thus the validation
246	 * step below suffices.
247	 */
248	smp_mb(); /* D */
249
250	return srcu_readers_seq_idx(sp, idx) == seq;
251}
252
253/**
254 * srcu_readers_active - returns approximate number of readers.
255 * @sp: which srcu_struct to count active readers (holding srcu_read_lock).
256 *
257 * Note that this is not an atomic primitive, and can therefore suffer
258 * severe errors when invoked on an active srcu_struct.  That said, it
259 * can be useful as an error check at cleanup time.
260 */
261static int srcu_readers_active(struct srcu_struct *sp)
262{
263	int cpu;
264	unsigned long sum = 0;
265
266	for_each_possible_cpu(cpu) {
267		sum += ACCESS_ONCE(per_cpu_ptr(sp->per_cpu_ref, cpu)->c[0]);
268		sum += ACCESS_ONCE(per_cpu_ptr(sp->per_cpu_ref, cpu)->c[1]);
269	}
270	return sum;
271}
272
273/**
274 * cleanup_srcu_struct - deconstruct a sleep-RCU structure
275 * @sp: structure to clean up.
276 *
277 * Must invoke this after you are finished using a given srcu_struct that
278 * was initialized via init_srcu_struct(), else you leak memory.
279 */
280void cleanup_srcu_struct(struct srcu_struct *sp)
281{
282	int sum;
283
284	sum = srcu_readers_active(sp);
285	WARN_ON(sum);  /* Leakage unless caller handles error. */
286	if (sum != 0)
287		return;
288	free_percpu(sp->per_cpu_ref);
289	sp->per_cpu_ref = NULL;
290}
291EXPORT_SYMBOL_GPL(cleanup_srcu_struct);
292
293/*
294 * Counts the new reader in the appropriate per-CPU element of the
295 * srcu_struct.  Must be called from process context.
296 * Returns an index that must be passed to the matching srcu_read_unlock().
297 */
298int __srcu_read_lock(struct srcu_struct *sp)
299{
300	int idx;
301
302	preempt_disable();
303	idx = rcu_dereference_index_check(sp->completed,
304					  rcu_read_lock_sched_held()) & 0x1;
305	ACCESS_ONCE(this_cpu_ptr(sp->per_cpu_ref)->c[idx]) += 1;
306	smp_mb(); /* B */  /* Avoid leaking the critical section. */
307	ACCESS_ONCE(this_cpu_ptr(sp->per_cpu_ref)->seq[idx]) += 1;
308	preempt_enable();
309	return idx;
310}
311EXPORT_SYMBOL_GPL(__srcu_read_lock);
312
313/*
314 * Removes the count for the old reader from the appropriate per-CPU
315 * element of the srcu_struct.  Note that this may well be a different
316 * CPU than that which was incremented by the corresponding srcu_read_lock().
317 * Must be called from process context.
318 */
319void __srcu_read_unlock(struct srcu_struct *sp, int idx)
320{
321	preempt_disable();
322	smp_mb(); /* C */  /* Avoid leaking the critical section. */
323	ACCESS_ONCE(this_cpu_ptr(sp->per_cpu_ref)->c[idx]) -= 1;
324	preempt_enable();
325}
326EXPORT_SYMBOL_GPL(__srcu_read_unlock);
327
328/*
329 * We use an adaptive strategy for synchronize_srcu() and especially for
330 * synchronize_srcu_expedited().  We spin for a fixed time period
331 * (defined below) to allow SRCU readers to exit their read-side critical
332 * sections.  If there are still some readers after 10 microseconds,
333 * we repeatedly block for 1-millisecond time periods.  This approach
334 * has done well in testing, so there is no need for a config parameter.
335 */
336#define SRCU_RETRY_CHECK_DELAY		5
337#define SYNCHRONIZE_SRCU_TRYCOUNT	2
338#define SYNCHRONIZE_SRCU_EXP_TRYCOUNT	12
339
340/*
341 * @@@ Wait until all pre-existing readers complete.  Such readers
342 * will have used the index specified by "idx".
343 * the caller should ensures the ->completed is not changed while checking
344 * and idx = (->completed & 1) ^ 1
345 */
346static bool try_check_zero(struct srcu_struct *sp, int idx, int trycount)
347{
348	for (;;) {
349		if (srcu_readers_active_idx_check(sp, idx))
350			return true;
351		if (--trycount <= 0)
352			return false;
353		udelay(SRCU_RETRY_CHECK_DELAY);
354	}
355}
356
357/*
358 * Increment the ->completed counter so that future SRCU readers will
359 * use the other rank of the ->c[] and ->seq[] arrays.  This allows
360 * us to wait for pre-existing readers in a starvation-free manner.
361 */
362static void srcu_flip(struct srcu_struct *sp)
363{
364	sp->completed++;
365}
366
367/*
368 * Enqueue an SRCU callback on the specified srcu_struct structure,
369 * initiating grace-period processing if it is not already running.
370 */
371void call_srcu(struct srcu_struct *sp, struct rcu_head *head,
372		void (*func)(struct rcu_head *head))
373{
374	unsigned long flags;
375
376	head->next = NULL;
377	head->func = func;
378	spin_lock_irqsave(&sp->queue_lock, flags);
379	rcu_batch_queue(&sp->batch_queue, head);
380	if (!sp->running) {
381		sp->running = true;
382		queue_delayed_work(system_nrt_wq, &sp->work, 0);
383	}
384	spin_unlock_irqrestore(&sp->queue_lock, flags);
385}
386EXPORT_SYMBOL_GPL(call_srcu);
387
388struct rcu_synchronize {
389	struct rcu_head head;
390	struct completion completion;
391};
392
393/*
394 * Awaken the corresponding synchronize_srcu() instance now that a
395 * grace period has elapsed.
396 */
397static void wakeme_after_rcu(struct rcu_head *head)
398{
399	struct rcu_synchronize *rcu;
400
401	rcu = container_of(head, struct rcu_synchronize, head);
402	complete(&rcu->completion);
403}
404
405static void srcu_advance_batches(struct srcu_struct *sp, int trycount);
406static void srcu_reschedule(struct srcu_struct *sp);
407
408/*
409 * Helper function for synchronize_srcu() and synchronize_srcu_expedited().
410 */
411static void __synchronize_srcu(struct srcu_struct *sp, int trycount)
412{
413	struct rcu_synchronize rcu;
414	struct rcu_head *head = &rcu.head;
415	bool done = false;
416
417	rcu_lockdep_assert(!lock_is_held(&sp->dep_map) &&
418			   !lock_is_held(&rcu_bh_lock_map) &&
419			   !lock_is_held(&rcu_lock_map) &&
420			   !lock_is_held(&rcu_sched_lock_map),
421			   "Illegal synchronize_srcu() in same-type SRCU (or RCU) read-side critical section");
422
423	init_completion(&rcu.completion);
424
425	head->next = NULL;
426	head->func = wakeme_after_rcu;
427	spin_lock_irq(&sp->queue_lock);
428	if (!sp->running) {
429		/* steal the processing owner */
430		sp->running = true;
431		rcu_batch_queue(&sp->batch_check0, head);
432		spin_unlock_irq(&sp->queue_lock);
433
434		srcu_advance_batches(sp, trycount);
435		if (!rcu_batch_empty(&sp->batch_done)) {
436			BUG_ON(sp->batch_done.head != head);
437			rcu_batch_dequeue(&sp->batch_done);
438			done = true;
439		}
440		/* give the processing owner to work_struct */
441		srcu_reschedule(sp);
442	} else {
443		rcu_batch_queue(&sp->batch_queue, head);
444		spin_unlock_irq(&sp->queue_lock);
445	}
446
447	if (!done)
448		wait_for_completion(&rcu.completion);
449}
450
451/**
452 * synchronize_srcu - wait for prior SRCU read-side critical-section completion
453 * @sp: srcu_struct with which to synchronize.
454 *
455 * Flip the completed counter, and wait for the old count to drain to zero.
456 * As with classic RCU, the updater must use some separate means of
457 * synchronizing concurrent updates.  Can block; must be called from
458 * process context.
459 *
460 * Note that it is illegal to call synchronize_srcu() from the corresponding
461 * SRCU read-side critical section; doing so will result in deadlock.
462 * However, it is perfectly legal to call synchronize_srcu() on one
463 * srcu_struct from some other srcu_struct's read-side critical section.
464 */
465void synchronize_srcu(struct srcu_struct *sp)
466{
467	__synchronize_srcu(sp, SYNCHRONIZE_SRCU_TRYCOUNT);
468}
469EXPORT_SYMBOL_GPL(synchronize_srcu);
470
471/**
472 * synchronize_srcu_expedited - Brute-force SRCU grace period
473 * @sp: srcu_struct with which to synchronize.
474 *
475 * Wait for an SRCU grace period to elapse, but be more aggressive about
476 * spinning rather than blocking when waiting.
477 *
478 * Note that it is illegal to call this function while holding any lock
479 * that is acquired by a CPU-hotplug notifier.  It is also illegal to call
480 * synchronize_srcu_expedited() from the corresponding SRCU read-side
481 * critical section; doing so will result in deadlock.  However, it is
482 * perfectly legal to call synchronize_srcu_expedited() on one srcu_struct
483 * from some other srcu_struct's read-side critical section, as long as
484 * the resulting graph of srcu_structs is acyclic.
485 */
486void synchronize_srcu_expedited(struct srcu_struct *sp)
487{
488	__synchronize_srcu(sp, SYNCHRONIZE_SRCU_EXP_TRYCOUNT);
489}
490EXPORT_SYMBOL_GPL(synchronize_srcu_expedited);
491
492/**
493 * srcu_barrier - Wait until all in-flight call_srcu() callbacks complete.
494 */
495void srcu_barrier(struct srcu_struct *sp)
496{
497	synchronize_srcu(sp);
498}
499EXPORT_SYMBOL_GPL(srcu_barrier);
500
501/**
502 * srcu_batches_completed - return batches completed.
503 * @sp: srcu_struct on which to report batch completion.
504 *
505 * Report the number of batches, correlated with, but not necessarily
506 * precisely the same as, the number of grace periods that have elapsed.
507 */
508long srcu_batches_completed(struct srcu_struct *sp)
509{
510	return sp->completed;
511}
512EXPORT_SYMBOL_GPL(srcu_batches_completed);
513
514#define SRCU_CALLBACK_BATCH	10
515#define SRCU_INTERVAL		1
516
517/*
518 * Move any new SRCU callbacks to the first stage of the SRCU grace
519 * period pipeline.
520 */
521static void srcu_collect_new(struct srcu_struct *sp)
522{
523	if (!rcu_batch_empty(&sp->batch_queue)) {
524		spin_lock_irq(&sp->queue_lock);
525		rcu_batch_move(&sp->batch_check0, &sp->batch_queue);
526		spin_unlock_irq(&sp->queue_lock);
527	}
528}
529
530/*
531 * Core SRCU state machine.  Advance callbacks from ->batch_check0 to
532 * ->batch_check1 and then to ->batch_done as readers drain.
533 */
534static void srcu_advance_batches(struct srcu_struct *sp, int trycount)
535{
536	int idx = 1 ^ (sp->completed & 1);
537
538	/*
539	 * Because readers might be delayed for an extended period after
540	 * fetching ->completed for their index, at any point in time there
541	 * might well be readers using both idx=0 and idx=1.  We therefore
542	 * need to wait for readers to clear from both index values before
543	 * invoking a callback.
544	 */
545
546	if (rcu_batch_empty(&sp->batch_check0) &&
547	    rcu_batch_empty(&sp->batch_check1))
548		return; /* no callbacks need to be advanced */
549
550	if (!try_check_zero(sp, idx, trycount))
551		return; /* failed to advance, will try after SRCU_INTERVAL */
552
553	/*
554	 * The callbacks in ->batch_check1 have already done with their
555	 * first zero check and flip back when they were enqueued on
556	 * ->batch_check0 in a previous invocation of srcu_advance_batches().
557	 * (Presumably try_check_zero() returned false during that
558	 * invocation, leaving the callbacks stranded on ->batch_check1.)
559	 * They are therefore ready to invoke, so move them to ->batch_done.
560	 */
561	rcu_batch_move(&sp->batch_done, &sp->batch_check1);
562
563	if (rcu_batch_empty(&sp->batch_check0))
564		return; /* no callbacks need to be advanced */
565	srcu_flip(sp);
566
567	/*
568	 * The callbacks in ->batch_check0 just finished their
569	 * first check zero and flip, so move them to ->batch_check1
570	 * for future checking on the other idx.
571	 */
572	rcu_batch_move(&sp->batch_check1, &sp->batch_check0);
573
574	/*
575	 * SRCU read-side critical sections are normally short, so check
576	 * at least twice in quick succession after a flip.
577	 */
578	trycount = trycount < 2 ? 2 : trycount;
579	if (!try_check_zero(sp, idx^1, trycount))
580		return; /* failed to advance, will try after SRCU_INTERVAL */
581
582	/*
583	 * The callbacks in ->batch_check1 have now waited for all
584	 * pre-existing readers using both idx values.  They are therefore
585	 * ready to invoke, so move them to ->batch_done.
586	 */
587	rcu_batch_move(&sp->batch_done, &sp->batch_check1);
588}
589
590/*
591 * Invoke a limited number of SRCU callbacks that have passed through
592 * their grace period.  If there are more to do, SRCU will reschedule
593 * the workqueue.
594 */
595static void srcu_invoke_callbacks(struct srcu_struct *sp)
596{
597	int i;
598	struct rcu_head *head;
599
600	for (i = 0; i < SRCU_CALLBACK_BATCH; i++) {
601		head = rcu_batch_dequeue(&sp->batch_done);
602		if (!head)
603			break;
604		local_bh_disable();
605		head->func(head);
606		local_bh_enable();
607	}
608}
609
610/*
611 * Finished one round of SRCU grace period.  Start another if there are
612 * more SRCU callbacks queued, otherwise put SRCU into not-running state.
613 */
614static void srcu_reschedule(struct srcu_struct *sp)
615{
616	bool pending = true;
617
618	if (rcu_batch_empty(&sp->batch_done) &&
619	    rcu_batch_empty(&sp->batch_check1) &&
620	    rcu_batch_empty(&sp->batch_check0) &&
621	    rcu_batch_empty(&sp->batch_queue)) {
622		spin_lock_irq(&sp->queue_lock);
623		if (rcu_batch_empty(&sp->batch_done) &&
624		    rcu_batch_empty(&sp->batch_check1) &&
625		    rcu_batch_empty(&sp->batch_check0) &&
626		    rcu_batch_empty(&sp->batch_queue)) {
627			sp->running = false;
628			pending = false;
629		}
630		spin_unlock_irq(&sp->queue_lock);
631	}
632
633	if (pending)
634		queue_delayed_work(system_nrt_wq, &sp->work, SRCU_INTERVAL);
635}
636
637/*
638 * This is the work-queue function that handles SRCU grace periods.
639 */
640static void process_srcu(struct work_struct *work)
641{
642	struct srcu_struct *sp;
643
644	sp = container_of(work, struct srcu_struct, work.work);
645
646	srcu_collect_new(sp);
647	srcu_advance_batches(sp, 1);
648	srcu_invoke_callbacks(sp);
649	srcu_reschedule(sp);
650}